Original title: Benchmarking Observable Platform Dynamics Beyond Persistence in Three Open Educational Datasets
Hongkui Li, Xueting Liu, Mingliang Gao · September 2026
Educational theories about artificial intelligence often describe learning processes that online school websites simply do not track. Researchers wanted to know how well basic platform data can predict a student's future habits. They separated learning ideas from four trackable digital measures: total achievement, student engagement, the range of resources used, and overall digital activity. They tested models that use past behavior to forecast future behavior. They found that past behavior does predict future activity better than assuming a student never changes. However, nearly all the predictive power came from looking at that exact same activity's own history. Checking other behaviors barely helped. Predicting future achievement was especially weak. Because predictions varied heavily across platforms, researchers must carefully test what local data actually means first.
Imagine predicting how well a friend will do on next week's quiz. The simplest guess is to copy last week's result, and extra details like pages read or hours studied seem like they should help. But for the actual quiz score, those extra details barely change the guess, and sometimes they make it worse. In this paper, the researchers tested whether adding digital traces of student activity improves predictions in online courses, and found it helps for some measures but not for achievement.
Online platforms are increasingly used to study how students learn with AI, but the records they collect were not designed for that purpose. This paper shows that extra digital traces can predict behavior better than a simple baseline, yet for actual achievement they add almost nothing — and the results shift across datasets. So anyone using these logs to draw conclusions about learning must first check what the signals mean in their own context.
The abstract explicitly says the study did not test generative-AI effects or causal educational mechanisms, and formal measurement invariance was not assessed. Results were outcome- and dataset-specific, so no general conclusion about which platform signals matter. The authors call for local validation before using these indicators in practice.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Benchmarking Observable Platform Dynamics Beyond Persistence in Three Open Educational Datasets — Hongkui Li, Xueting Liu, Mingliang Gao — Education Sciences — doi:10.3390/educsci16091572
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.